Transformer Geometry Observatory TGO-II: Representational Similarity Observatory

Kaustubh Kapil; Kishor P. Upla

arxiv: 2607.02386 · v1 · pith:QGWPGDUZnew · submitted 2026-07-02 · 💻 cs.CV · cs.LG

Transformer Geometry Observatory TGO-II: Representational Similarity Observatory

Kaustubh Kapil , Kishor P. Upla This is my paper

Pith reviewed 2026-07-03 15:12 UTC · model grok-4.3

classification 💻 cs.CV cs.LG

keywords vision transformersrepresentational geometryCKASVCCAintrinsic dimensionalitytoken covariancetraining dynamicslayer specialization

0 comments

The pith

Vision Transformers increase representational complexity through progressively richer transformations while preserving strong token interaction structure during learning.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines how the internal geometry of representations in a Vision Transformer changes over the course of supervised training. It tracks four geometric measures across layers and finds that layer-to-layer similarity drops steadily, the local dimensionality of the representation space rises then levels off, and pairwise token relationships stay strongly coupled rather than loosening. These patterns together indicate that specialization and manifold expansion happen at the same time, without the need for tokens to become independent. A reader would care because the results point to a different source for growing model capacity than the common assumption of progressive token decoupling.

Core claim

Analysis of ViT-Small/16 with CKA, SVCCA, TwoNN-ID, and token covariance shows that CKA and SVCCA decrease throughout training, intrinsic dimensionality increases before stabilizing, and strong token interaction structure persists. These observations indicate that representation complexity and layer specialization emerge simultaneously, with manifold expansion occurring without token decoupling, so that complexity arises from progressively richer transformations while token coupling remains intact.

What carries the argument

TGO-II framework that applies CKA, SVCCA, TwoNN-ID, and token covariance metrics to measure layer specialization, manifold expansion, and token coupling across training epochs.

If this is right

Representational specialization across layers increases steadily as training proceeds.
The representation manifold expands to more local degrees of freedom before it stabilizes.
Strong token interactions are preserved rather than replaced by independence.
Complexity and specialization develop together through richer per-layer transformations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same geometric measures could be applied to language transformers to test whether token coupling persists there as well.
If the pattern holds, training methods that encourage richer transformations without forcing token independence might be more effective than current decoupling-focused approaches.

Load-bearing premise

The four chosen metrics accurately track the underlying geometric changes in the ViT-Small/16 representations without introducing their own artifacts.

What would settle it

Measuring a substantial drop in token covariance at the same time that intrinsic dimensionality is still rising would contradict the claim that manifold expansion occurs without token decoupling.

Figures

Figures reproduced from arXiv: 2607.02386 by Kaustubh Kapil, Kishor P. Upla.

**Figure 2.** Figure 2: a) Shows how the mean SVCCA shows a sudden drop and then stabilizes over the training schedule. b) Adjacent Layer [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Evolution of TwoNN intrinsic dimensionality. The left figure illustrates the evolution of intrinsic dimensionality across [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: Token Covariance matrces for different layers and epochs [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

read the original abstract

While Vision Transformers have achieved remarkable success across computer vision and language applications, the geometric evolution of their internal representations throughout training remains insufficiently understood. Existing analyses primarily focus on attention mechanisms and downstream performance, leaving the evolution of representation geometry largely unexplored. In this work, we present Transformer Geometry Observatory-II (TGO-II), a representation geometry analysis framework designed to investigate how Transformer representations evolve during supervised training. TGO-II analyzes Vision Transformer (ViT-Small/16) representations using Centered Kernel Alignment (CKA), Singular Vector Canonical Correlation Analysis (SVCCA), Two-Nearest Neighbor Intrinsic Dimensionality (TwoNN-ID), and token covariance analysis. Our experiments reveal three key observations. First, both CKA and SVCCA progressively decrease throughout training, indicating increasing representational specialization across Transformer layers. Second, intrinsic dimensionality consistently increases before stabilizing, suggesting progressive expansion of the representation manifold into a larger set of locally accessible degrees of freedom. Third, token covariance and coupling analyses demonstrate that strong token interaction structure persists throughout training, challenging the hypothesis that increasing representational complexity arises primarily from progressive token independence. These findings suggest that representation complexity and layer specialization emerge simultaneously during training. Manifold expansion appears to occur without token decoupling. Together, these observations motivate a new hypothesis in which Vision Transformers increase representational complexity through progressively richer transformations while preserving strong token interaction structure during learning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper applies four standard metrics to one ViT-Small run and reports three training trends, but supplies almost no methods or controls in the abstract.

read the letter

The core of this work is an observational run on ViT-Small/16 under supervised training. They track CKA and SVCCA between layers, TwoNN intrinsic dimension, and token covariance, then note that layer similarity falls, intrinsic dimension rises then levels off, and token covariance stays high. From that they suggest complexity grows via richer transformations without token decoupling.

The paper does one thing cleanly: it packages four off-the-shelf tools into a single training-time monitor and uses them to push back on the token-independence story. That combination is not a new derivation, but it is a practical way to look at the same model from multiple geometric angles at once.

The soft spot is obvious from the abstract alone. No protocol, no hyperparameter controls, no error bars, no ablations, and no mention of multiple seeds or different data regimes. Without those, it is impossible to tell whether the three trends are stable or sensitive to the exact training setup. The stress-test note is right that there is no circularity in the claims, but that does not make up for the missing experimental detail.

This is for people who already work on representation analysis in transformers and want a ready-made set of diagnostics to try on their own runs. A reader who needs new theory or broad claims will not get much. It is narrow in scope and incremental in method.

I would bring the full paper to a reading group to see whether the methods section actually delivers the missing controls. I would not cite it in the next year on current evidence. It is worth sending to referees if the full manuscript shows reproducible code and basic robustness checks; otherwise it stays preliminary.

Referee Report

2 major / 2 minor

Summary. The paper introduces the Transformer Geometry Observatory-II (TGO-II) framework to study the geometric evolution of representations in Vision Transformers during supervised training. It applies CKA, SVCCA, TwoNN-ID, and token covariance metrics to a ViT-Small/16 model and reports three observations: progressive decreases in CKA and SVCCA indicating increasing representational specialization across layers; intrinsic dimensionality that increases before stabilizing, suggesting manifold expansion; and persistent token covariance and coupling, indicating that strong token interaction structure is maintained. These motivate the hypothesis that Vision Transformers increase representational complexity through progressively richer transformations while preserving strong token interaction structure.

Significance. If the reported trends are robust, the work supplies concrete empirical observations on ViT representation geometry that could guide architecture design and training analysis. The multi-metric approach (similarity, intrinsic dimension, and covariance) is a strength relative to single-metric studies. The single-model, single-regime scope, however, limits immediate generalizability.

major comments (2)

[Experiments section] Experiments / Methods section: the manuscript supplies no description of the training protocol (dataset, optimizer, schedule, number of epochs, or random seeds), the precise layer-wise and epoch-wise computation of the four metrics, or any ablation on hyperparameter sensitivity. These details are load-bearing because the three stated observations are direct empirical claims about metric trajectories.
[Results section] Results section: no error bars, multiple independent runs, or statistical tests accompany the reported trends (e.g., “progressively decrease,” “consistently increases before stabilizing”). Without them the reliability of the specialization and manifold-expansion claims cannot be assessed.

minor comments (2)

The abstract would benefit from a one-sentence statement of the model size, patch size, and dataset to orient readers immediately.
Consider adding a single composite figure that overlays the four metric trajectories across training epochs for visual clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. The two major comments correctly identify gaps in experimental documentation and statistical rigor that limit the strength of the empirical claims. We will revise the manuscript to address both points fully.

read point-by-point responses

Referee: [Experiments section] Experiments / Methods section: the manuscript supplies no description of the training protocol (dataset, optimizer, schedule, number of epochs, or random seeds), the precise layer-wise and epoch-wise computation of the four metrics, or any ablation on hyperparameter sensitivity. These details are load-bearing because the three stated observations are direct empirical claims about metric trajectories.

Authors: We agree that these details are essential. In the revised manuscript we will add a complete description of the supervised training protocol (ImageNet-1k, ViT-Small/16, optimizer, schedule, epochs, and seeds) together with the exact layer-wise and epoch-wise procedures used to compute CKA, SVCCA, TwoNN-ID, and token covariance. We will also include a short hyperparameter-sensitivity discussion focused on the metrics that drive the three main observations. revision: yes
Referee: [Results section] Results section: no error bars, multiple independent runs, or statistical tests accompany the reported trends (e.g., “progressively decrease,” “consistently increases before stabilizing”). Without them the reliability of the specialization and manifold-expansion claims cannot be assessed.

Authors: We accept this criticism. The revised Results section will report trends aggregated over multiple independent training runs, include error bars (standard deviation across seeds), and add appropriate statistical tests for the reported monotonic or stabilizing behaviors. revision: yes

Circularity Check

0 steps flagged

No significant circularity: purely observational empirical study

full rationale

The paper applies standard, externally defined metrics (CKA, SVCCA, TwoNN-ID, token covariance) to track representation evolution in a single ViT-Small/16 model under supervised training. All reported trends are direct outputs of these measurements on the model's activations; no equations, fitted parameters, or predictions are derived from quantities defined inside the paper itself. No self-citation chains, ansatzes, or uniqueness theorems are invoked to support any derivation. The central claims are therefore observational hypotheses conditioned on accepting the metrics at face value, with no reduction of outputs to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that the four listed metrics are faithful probes of representation geometry; no free parameters are fitted, no new entities are postulated, and the work is purely observational.

axioms (1)

domain assumption Centered Kernel Alignment, Singular Vector Canonical Correlation Analysis, Two-Nearest Neighbor Intrinsic Dimensionality, and token covariance are appropriate and sufficient measures for characterizing the geometric evolution of Transformer representations.
The framework description invokes these metrics to support all three key observations and the final hypothesis.

pith-pipeline@v0.9.1-grok · 5772 in / 1316 out tokens · 26924 ms · 2026-07-03T15:12:39.782142+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

7 extracted references · 7 canonical work pages · 3 internal anchors

[1]

Transformer Geometry Observatory TGO-I: Spectral Geometry Observatory

K. Kapil and K. P. Upla, “Transformer geometry observatory tgo-i: Spectral geometry observatory,” 2026. [Online]. Available: https://arxiv.org/abs/2606.19249

work page internal anchor Pith review Pith/arXiv arXiv 2026
[2]

Similarity of Neural Network Representations Revisited

S. Kornblith, M. Norouzi, H. Lee, and G. Hinton, “Similarity of neural network representations revisited,” 2019. [Online]. Available: https://arxiv.org/abs/1905.00414

work page internal anchor Pith review Pith/arXiv arXiv 2019
[3]

Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability,

M. Raghu, J. Gilmer, J. Yosinski, and J. Sohl-Dickstein, “Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability,” 2017. [Online]. Available: https://arxiv.org/abs/1706. 05806

work page 2017
[4]

The training accuracy of two-layer neural networks: its estimation and understanding using random datasets,

S. Guan and M. Loew, “The training accuracy of two-layer neural networks: its estimation and understanding using random datasets,”

work page
[5]

Available: https://arxiv.org/abs/2010.13380

[Online]. Available: https://arxiv.org/abs/2010.13380

work page arXiv 2010
[6]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” 2021. [Online]. Available: https://arxiv.org/abs/2010.11929

work page internal anchor Pith review Pith/arXiv arXiv 2021
[7]

Imagenet: A large-scale hierarchical image database,

J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009, pp. 248–255

work page 2009

[1] [1]

Transformer Geometry Observatory TGO-I: Spectral Geometry Observatory

K. Kapil and K. P. Upla, “Transformer geometry observatory tgo-i: Spectral geometry observatory,” 2026. [Online]. Available: https://arxiv.org/abs/2606.19249

work page internal anchor Pith review Pith/arXiv arXiv 2026

[2] [2]

Similarity of Neural Network Representations Revisited

S. Kornblith, M. Norouzi, H. Lee, and G. Hinton, “Similarity of neural network representations revisited,” 2019. [Online]. Available: https://arxiv.org/abs/1905.00414

work page internal anchor Pith review Pith/arXiv arXiv 2019

[3] [3]

Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability,

M. Raghu, J. Gilmer, J. Yosinski, and J. Sohl-Dickstein, “Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability,” 2017. [Online]. Available: https://arxiv.org/abs/1706. 05806

work page 2017

[4] [4]

The training accuracy of two-layer neural networks: its estimation and understanding using random datasets,

S. Guan and M. Loew, “The training accuracy of two-layer neural networks: its estimation and understanding using random datasets,”

work page

[5] [5]

Available: https://arxiv.org/abs/2010.13380

[Online]. Available: https://arxiv.org/abs/2010.13380

work page arXiv 2010

[6] [6]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” 2021. [Online]. Available: https://arxiv.org/abs/2010.11929

work page internal anchor Pith review Pith/arXiv arXiv 2021

[7] [7]

Imagenet: A large-scale hierarchical image database,

J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009, pp. 248–255

work page 2009